Pith. sign in

REVIEW 5 major objections 4 minor 29 references

Score-based Diffusion Model for Unpaired Virtual Histology Staining

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that an MI-guided score-based diffusion model can translate unpaired H&E histology images into virtual IHC stains with better structural fidelity, color matching, and cellular-level consistency than existing translation…

desk verdict A plausible compositional diffusion pipeline for unpaired virtual staining with moderate gains, but the cross-domain MI estimator and missing validation keep the central claim unproven. read the letter →

arxiv 2506.23184 v1 pith:ERU45R2Z submitted 2025-06-29 eess.IV cs.AIcs.CV

Anran Liu , Xiaofei Wang , Jing Cai , Chao Li This is my paper
classification eess.IVcs.AIcs.CV
keywords virtualstainingH&EtoIHCtranslationunpairedimagescore-baseddiffusionmodelmutualinformationestimationenergy-guidedreverseSDEcontrastivelearninghistopathologyimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to show that unpaired virtual staining from H&E to IHC can be done by a score-based diffusion model guided by mutual information, without needing pixel-aligned training pairs. If correct, it offers a way to generate IHC-like stains from routine H&E slides, which would matter for clinical diagnosis when tissue or antibodies are scarce. The method reports better information fidelity, color matching, and cellular-level positive-signal consistency than prior translation models on two public datasets.

What carries the argument

The load-bearing object is the global MI estimator, trained on IHC image-gradient pairs with a mutual-information neural estimation lower bound, whose output approximates the mutual information between the source gradient map and the generated image. The reverse process is steered by the energy function $M(y,x,t)=\lambda_t U(y,x,t)+(1-\lambda_t)I(x;y)$, where $U(y)=H(y)-I(x;y)$ is the unique staining information and $\lambda_t=t/S$, so early timesteps emphasize staining and later timesteps emphasize structure; solving the adapted SDE reduces to shifting the mean of the pretrained diffusion kernel by the gradient of this energy. Local consistency is added by a patch-wise contrastive loss over encoder features of positive and negative patches.

What would settle it

Compute the MI estimator's scores on matched H&E and IHC gradient patches from the same tissue sections: if the estimator does not rank corresponding structures similarly across the two stain domains, the cross-domain guidance is misfounded and the reported fidelity improvements would not generalize to new tissues.

Watch

Extended reading notes

Core claim

MIU-Diff is a two-stage score-based diffusion framework. Stage one learns the marginal distribution of IHC images with an unconditional stochastic differential equation; stage two runs a reverse SDE whose transition-kernel mean is shifted by an MI-guided energy function that balances unique staining information against shared structural information, with a timestep-dependent weight. A local MI-driven contrastive loss updates the score model after an intermediate timestep to enforce cellular-level structural consistency. The paper claims this pipeline outperforms GAN- and diffusion-based baselines on the BCI and MIST datasets in visual information fidelity, histogram correlation, and integrated optical density deviation, while noting that higher PSNR alone is not meaningful without pixel alignment.

Load-bearing premise

The estimator that measures shared structure is trained only on IHC image-gradient pairs, yet during generation it is fed an H&E-derived gradient map; the paper gives no direct evidence that gradient features of H&E and IHC are comparable enough for the estimate to be valid in the reverse process.

Editorial extensions

If this is right

  • Unpaired H&E-IHC pairs are sufficient; pixel-aligned registration is not required, so the method fits clinical slides cut from consecutive sections.
  • The timestep-dependent trade-off between staining and structure provides a single control knob for adjusting staining intensity across different tissue types and protein markers.
  • Local patch contrastive learning enforces cellular-level structural consistency, which is the clinically relevant measure for trusting positive-signal regions.
  • Ablations show that removing either the MI guidance or the contrastive loss degrades fidelity, so both components carry the reported improvement.
  • The approach outperforms GAN-based (CUT, ASP) and diffusion-based (ILVR, SDEdit, EGSDE) baselines on BCI-HER2, MIST-HER2, and MIST-Ki67.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the cross-domain MI transfer holds, the same estimator-plus-energy recipe could apply to other stain pairs or to unpaired translation tasks beyond histology, since the energy term never uses explicit domain labels.
  • A reader could probe the estimator on H&E gradients versus IHC gradients of the same tissue to check whether the assumed cross-domain comparability actually holds; the paper does not report such a check.
  • The IOD deviation metric measures segmented positive regions, so the reported gains suggest the method may be usable for receptor status scoring, but clinical validation on whole-slide images with variable preprocessing would be the next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The manuscript proposes MIU-Diff, a two-stage score-based diffusion model for unpaired H&E-to-IHC virtual staining. The method trains a marginal diffusion model on IHC images and then guides the reverse process with a mutual-information (MI) energy function intended to disentangle staining-specific information U(y) from shared structural information I(x;y). A global MI estimator is trained on IHC image/gradient pairs, and a local MI-driven contrastive loss is used to promote cellular-level structural consistency. Experiments on the BCI and MIST datasets compare against two GAN-based and three diffusion-based baselines using PSNR, VIF, histogram correlation, and IOD deviation, with additional ablations and hyperparameter sensitivity analysis.

Significance. If the empirical claims hold, the paper makes a useful contribution to unpaired virtual staining by integrating MI-guided energy into a score-based diffusion framework and demonstrating gains in VIF, histogram correlation, and IOD deviation on two public datasets. The paper is not circular in the sense of fitting a constant: the MI estimator is trained on separate IHC pairs and the hyperparameters are tuned on validation sets. The strengths include the use of two public benchmarks, comparison with multiple state-of-the-art methods, and ablations of the main components. However, the central claims currently rest on an ambiguous energy formulation, an unvalidated cross-domain MI estimator, an unclear training-time versus per-image optimization procedure, and a lack of statistical validation.

major comments (5)
  1. [Section 2.2, Eq. (2)] Equation (2) first defines M(y,x,t) = λ_t U(y,x,t) + (1−λ_t) I(y,x,t), but the second equality rewrites it as −λ_t E_{q_{t|0}}U − (1−λ_t) E_{q_{t|0}}I, which has opposite signs. Since Eq. (1) uses −∇_y M, the actual direction of the guidance is ambiguous: if the second form is intended, the reverse SDE performs gradient ascent on U and I; if the first form is intended, it performs gradient descent. The accompanying text says the energy prioritizes staining details before shared structural information, but with λ_t = t/S the weighting changes in a way that a literal reading does not make consistent. Please clarify the intended sign convention and explicitly define the t-dependence of U(y,x,t) and I(y,x,t).
  2. [Section 2.2, MI estimator] The global MI estimator G_θ is trained only on IHC image and IHC gradient-map pairs, as shown in Fig. 2(a). At inference, however, it is applied to H&E gradient maps g_{x_t} paired with the generated y_t in order to compute I(x_t;y_t). The statement that the MI 'can be simply obtained using the well-trained MI estimator' is not supported by any evidence that G_θ's joint-versus-marginal decision rule transfers across stain domains. Because a miscalibrated estimate enters ∇_y M in Eq. (1), it can directly perturb the reverse trajectory and distort either staining or structure. Please add a validation of this cross-domain transfer, for example by training G_θ on H&E pairs and comparing the resulting energy guidance, or by reporting MI estimates on held-out same-domain and cross-domain pairs.
  3. [Section 2.3 and Table 2] The manuscript does not specify whether the local MI-driven contrastive loss ℓ_PCL is minimized on the training set or is applied as per-image test-time optimization. Section 2.3 says that 'from t0 onward, we update the score-based model e(y,t) using a patch-wise contrastive loss,' and the implementation details only report N=300 and t0=40. If ℓ_PCL is optimized separately for each test image during the reverse process, the comparison against one-shot baselines (ILVR, SDEdit, EGSDE, CUT, ASP) is not controlled for computational budget, and the method is not a standard unpaired translation model. Please state clearly which setting is used, and if test-time optimization is performed, report the additional cost and compare against baselines under an equal budget.
  4. [Tables 1 and 2] All quantitative claims are based on single point estimates without error bars or significance tests. The reported differences are sometimes very small (e.g., VIF 0.611 vs 0.610 for MIST-HER2 in Table 1), and diffusion sampling is stochastic, so the assertion that MIU-Diff 'achieves the best performance' is not yet statistically supported. Please report means and standard deviations over at least three random seeds and perform paired significance tests (or an equivalent) on the test sets.
  5. [Section 2.2, U(y)=H(y)-I(x;y)] The decomposition U(y_t) = H(y_t) − I(x_t;y_t), described as the 'unique staining information' in y_t, is asserted without proof. In information theory, H(Y)−I(X;Y)=H(Y|X), so this equality identifies staining information with the entropy of the generated image conditioned on the color-removed H&E source. This is only meaningful if x_t captures all shared structural information and if the residual conditional entropy corresponds exactly to staining content. Since this decomposition underlies the disentanglement claim and the energy guidance in Eq. (2), please provide a formal definition or an empirical validation that U is sensitive to staining changes but not to structural changes.
minor comments (4)
  1. [Abstract and Introduction] There are several typographical errors, including 'lacksspecificity' in the abstract, 'the our superiority' in the abstract, and 'unoque' in the Introduction.
  2. [Section 2.2 vs Section 3] The total number of denoising timesteps is defined as N in Section 2.2 but is called T in Table 2 and in the implementation text; please use a single notation consistently.
  3. [Section 2.3] The patch-wise contrastive loss is not fully specified: the number of sampled patches, the encoder layers used, feature normalization, and the temperature in the cross-entropy loss are not reported, which prevents reproduction.
  4. [Section 2.2, Fig. 2(b)] In Eq. (2) and Fig. 2(b), the expectations E_{q_{t|0}(x_t|x)} are not consistently written with their conditioning variable; clarify whether the energy is a function of the noised source x_t or the original x_0.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the MIU-Diff derivation is an empirical pipeline with external benchmarks; the cross-domain use of the MI estimator is an unvalidated transfer assumption, not a circular reduction.

full rationale

I examined the derivation chain from the two-stage score-based diffusion framework through the MI-guided energy function and the local contrastive loss. No step equates a predicted quantity to a fitted parameter by construction. The global MI estimator G_theta is trained to estimate I(g_y; y) on IHC image/gradient pairs (Section 2.2), and is then applied at inference to gradient maps of the source H&E image. This is an out-of-distribution extrapolation that may fail, but it is not circular: the paper never defines the H&E-gradient MI as equal to the IHC-trained objective, nor does it fit G_theta to the final H&E-to-IHC outputs. The decomposition U(y) = H(y) - I(x; y) is a definitional modeling choice, not a result derived from the data, so it does not smuggle the target into the inputs. The timestep-adaptive reverse SDE shifts the pretrained transition kernel by the gradient of the energy function; this is a standard guided-diffusion construction and does not reduce to the fitted MI estimator by identity. The central performance claim is supported by comparisons against external baselines (CUT, ASP, ILVR, SDEdit, EGSDE) on two public datasets, with hyperparameters selected on validation sets and then evaluated on test splits; this is ordinary hyperparameter tuning, not a fitted input renamed as a prediction. The only self-citation I found, reference [22] (Wang, Huang, Price, Li, MICCAI 2024), is an author prior work on spatial transcriptomics and is not load-bearing for any claim in this paper. No uniqueness theorem or ansatz is imported from the authors' own prior work. Accordingly, the appropriate finding is no significant circularity, score 0; the cross-domain MI-estimator concern is a correctness risk, not evidence of circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Most of the paper's machinery is composed of existing components such as score-based SDEs, MINE, PatchNCE, and EGSDE. The new part is the choice to combine them and the heuristic decomposition of entropy and mutual information into structure and staining. The main unpaid assumptions are cross-domain MI estimation and the decomposition of U(y).

free parameters (4)
  • Total denoising timesteps N = 300
    Chosen by validation; ablation shows N=100 and N=500 degrade balance of staining and structure (Table 2).
  • Contrastive learning start timestep t'_0 = 40
    Chosen by validation; Table 2 shows t'_0=80 causes sharpening and staining bias, while t'_0=10 reduces consistency.
  • Patch count and patch size for MI estimator
    The MI estimator and contrastive loss depend on patch sampling (L patches in Section 2.2), but exact values are omitted.
  • Histogram binning for entropy H(y)
    Entropy estimate depends on histogram of pixel values; bin count is not stated.
assumptions (5)
  • domain assumption The gradient map captures the structural information of an image.
    Used in Section 2.2 to justify the global MI estimator; no derivation or validation for histology images.
  • domain assumption The MI estimator trained on IHC gradient-image pairs generalizes to H&E gradient-image pairs.
    Section 2.2 applies the estimator to H&E gradient maps without evidence of cross-domain validity.
  • ad hoc to paper U(y_t)=H(y_t)-I(x_t;y_t) represents the unique staining information of y_t.
    Defined in Section 2.2 without proof that this decomposition isolates stain content from structure.
  • standard math The MI-guided energy function has lower curvature than the inverse covariance of the transition kernel.
    Needed for the Taylor expansion in Eq. (3), following the EGSDE derivation.
  • domain assumption Color removal from H&E images removes staining information while preserving structure.
    Section 2.1 removes color from x0 to obtain x'_0; this assumes H&E stain is not needed for the translation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Score-based Diffusion Model for Unpaired Virtual Histology Staining." pith.science (2026). https://pith.science/paper/ERU45R2Z

@misc{pith2026250623184,
  author       = {Pith},
  title        = {Pith review of: Score-based Diffusion Model for Unpaired Virtual Histology Staining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERU45R2Z}},
  note         = {Machine review of arXiv:2506.23184}
}
read the original abstract

Hematoxylin and eosin (H&E) staining visualizes histology but lacks specificity for diagnostic markers. Immunohistochemistry (IHC) staining provides protein-targeted staining but is restricted by tissue availability and antibody specificity. Virtual staining, i.e., computationally translating the H&E image to its IHC counterpart while preserving the tissue structure, is promising for efficient IHC generation. Existing virtual staining methods still face key challenges: 1) effective decomposition of staining style and tissue structure, 2) controllable staining process adaptable to diverse tissue and proteins, and 3) rigorous structural consistency modelling to handle the non-pixel-aligned nature of paired H&E and IHC images. This study proposes a mutual-information (MI)-guided score-based diffusion model for unpaired virtual staining. Specifically, we design 1) a global MI-guided energy function that disentangles the tissue structure and staining characteristics across modalities, 2) a novel timestep-customized reverse diffusion process for precise control of the staining intensity and structural reconstruction, and 3) a local MI-driven contrastive learning strategy to ensure the cellular level structural consistency between H&E-IHC images. Extensive experiments demonstrate the our superiority over state-of-the-art approaches, highlighting its biomedical potential. Codes will be open-sourced upon acceptance.

Figures

Figures reproduced from arXiv: 2506.23184 by the authors.

Figure 1
Figure 1. The overall workflow of MIU-Diff, including (a) disentanglement of the tissue structure and staining characteristics based on the MI-guided energy function, (b) controllable denoising process at each timestep with the proposed energy function, and (c) local MI-driven contrastive learning that ensures cross-modal structure consistency. the best of our knowledge, this is the first study for unpaired virtual histology … view at source ↗
Figure 2
Figure 2. Illustration of (a) the construction and training process of the global MI estima￾tor, and (b) the MI-guided energy function in the reverse diffusion, which is formulated by I(x; y) approximated using the MI estimator and U(y). pairs g 1 y , y 1  , . . . , g L y , y L  ∼ q(gy , y) as joint distribution, and L patch-level shuffled pairs g 1 y , y¯ 1  , . . . , g L y , y¯ L  ∼ q(gy )q(y) as marginal distribution, … view at source ↗
Figure 3
Figure 3. Visual comparisons on MISTKi67 and MISTHER2. trained for 40k iterations with batch size 4 and learning rate 1 × 10−4 using AdamW optimizer. The images are resized from 1024×1024 to 256×256 due to memory limitations. For validation and test, the batch size was 1, and the hyper-parameter N and t ′ 0 are set to 300 and 40, respectively. 3.1 Performance evaluation We compare our model with advanced image translation met… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages

  1. [1]

    arXiv preprint arXiv:1801.04062 (2018)

    Belghazi, M.I., Baratin, A., Rajeswar, S., Ozair, S., Bengio, Y., Courville, A., Hjelm, R.D.: Mine: mutual information neural estimation. arXiv preprint arXiv:1801.04062 (2018)

    arXiv 2018

  2. [2]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Chen, F., Zhang, R., Zheng, B., Sun, Y., He, J., Qin, W.: Pathological semantics- preserving learning for h&e-to-ihc virtual staining. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 384–394. Springer (2024)

    work page 2024

  3. [3]

    arXiv preprint arXiv:2108.02938 (2021)

    Choi, J., Kim, S., Jeong, Y., Gwon, Y., Yoon, S.: Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938 (2021)

    arXiv 2021

  4. [4]

    Cold spring harbor protocols2008(5), pdb–prot4986 (2008)

    Fischer, A.H., Jacobson, K.A., Rose, J., Zeller, R.: Hematoxylin and eosin staining of tissue and cell sections. Cold spring harbor protocols2008(5), pdb–prot4986 (2008)

    work page 2008

  5. [5]

    Journal of proteome research8(7), 3558–3567 (2009)

    Hanselmann, M., Kothe, U., Kirchner, M., Renard, B.Y., Amstalden, E.R., Glunde, K., Heeren, R.M., Hamprecht, F.A.: Toward digital staining using imaging mass spectrometry and random forests. Journal of proteome research8(7), 3558–3567 (2009)

    work page 2009

  6. [6]

    He, Y., Liu, Z., Qi, M., Ding, S., Zhang, P., Song, F., Ma, C., Wu, H., Cai, R., Feng, Y., et al.: Pst-diff: achieving high-consistency stain transfer by diffusion modelswithpathologicalandstructuralconstraints.IEEETransactionsonMedical Imaging (2024)

    work page 2024

  7. [7]

    arXiv preprint arXiv:2403.11340 (2024)

    Kataria, T., Knudsen, B., Elhabian, S.Y.: Staindiffuser: Multitask dual diffusion model for virtual staining. arXiv preprint arXiv:2403.11340 (2024)

    arXiv 2024

  8. [8]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Li, F., Hu, Z., Chen, W., Kak, A.: Adaptive supervised patchnce loss for learn- ing h&e-to-ihc stain translation with inconsistent groundtruth image pairs. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 632–641. Springer (2023)

    work page 2023

Show all 29 references
  1. [9]

    In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention

    Li, Y., Guan, X., Wang, Y., Zhang, Y.: Exploiting supervision information in weakly paired images for ihc virtual staining. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 113–122. Springer (2024)

    2024

  2. [10]

    International Conference on Medical Image Computing and Computer-Assisted Intervention pp

    Li F, Hu Z, C.W., A, K.: Adaptive supervised patchnce loss for learning h&e- to-ihc stain translation with inconsistent groundtruth image pairs. International Conference on Medical Image Computing and Computer-Assisted Intervention pp. 632–641 (2023)

    2023

  3. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, S., Zhu, C., Xu, F., Jia, X., Shi, Z., Jin, M.: Bci: Breast cancer immuno- histochemical image generation through pyramid pix2pix. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1815–1824 (2022)

    2022

  4. [12]

    IEEE transactions on medical imaging40(8), 1977–1989 (2021)

    Liu, S., Zhang, B., Liu, Y., Han, A., Shi, H., Guan, T., He, Y.: Unpaired stain transfer using pathology-consistent constrained generative adversarial networks. IEEE transactions on medical imaging40(8), 1977–1989 (2021)

    2021

  5. [13]

    IEEE Journal of Biomedical and Health Informatics (2024)

    Luo, Y., Yang, Q., Liu, Z., Shi, Z., Huang, W., Zheng, G., Cheng, J.: Target- guided diffusion models for unpaired cross-modality medical image translation. IEEE Journal of Biomedical and Health Informatics (2024)

    2024

  6. [14]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Ma, C., Rao, Y., Cheng, Y., Chen, C., Lu, J., Zhou, J.: Structure-preserving super resolution with gradient guidance. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7769–7778 (2020) Diffusion for Unpaired Virtual Histology Staining 11

    2020

  7. [15]

    Journal of Computational and Applied Mathematics290, 370–384 (2015)

    Mao, X.: The truncated euler–maruyama method for stochastic differential equa- tions. Journal of Computational and Applied Mathematics290, 370–384 (2015)

    2015

  8. [16]

    arXiv preprint arXiv:2108.01073 (2021)

    Meng, C., He, Y., Song, Y., Song, J., Wu, J., Zhu, J.Y., Ermon, S.: Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073 (2021)

    2021 arXiv

  9. [17]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Moghadam, P.A., Van Dalen, S., Martin, K.C., Lennerz, J., Yip, S., Farahani, H., Bashashati, A.: A morphology focused diffusion probabilistic model for synthesis of histopathology images. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp....

    2023

  10. [18]

    Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-imagetranslation.In:ComputerVision–ECCV2020:16thEuropeanCon- ference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 319–345. Springer (2020)

    2020

  11. [19]

    IEEE Transac- tions on image processing15(2), 430–444 (2006)

    Sheikh, H.R., Bovik, A.C.: Image information and visual quality. IEEE Transac- tions on image processing15(2), 430–444 (2006)

    2006

  12. [20]

    arXiv preprint arXiv:2011.13456 (2020)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

    2020 arXiv

  13. [21]

    arXiv preprint arXiv:2203.08382 (2022)

    Su, X., Song, J., Meng, C., Ermon, S.: Dual diffusion implicit bridges for image- to-image translation. arXiv preprint arXiv:2203.08382 (2022)

    2022 arXiv

  14. [22]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Wang, X., Huang, X., Price, S., Li, C.: Cross-modal diffusion modelling for super- resolved spatial transcriptomics. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 98–108. Springer (2024)

    2024

  15. [23]

    In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision

    Wang, Z., Zhao, L., Xing, W.: Stylediffusion: Controllable disentangled style trans- fer via diffusion models. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 7677–7689 (2023)

    2023

  16. [24]

    PloS one 16(10), e0258546 (2021)

    Wieslander, H., Gupta, A., Bergman, E., Hallström, E., Harrison, P.J.: Learning to see colours: Biologically relevant virtual staining for adipocyte cell images. PloS one 16(10), e0258546 (2021)

    2021

  17. [25]

    arXiv preprint arXiv:2302.03018 (2023)

    Xiang, T., Yurt, M., Syed, A.B., Setsompop, K., Chaudhari, A.: DDM2: Self- supervised diffusion mri denoising with generative diffusion models. arXiv preprint arXiv:2302.03018 (2023)

    2023 arXiv

  18. [26]

    Advances in Neural Information Processing Systems 36 (2024)

    Xu, S., Ma, Z., Huang, Y., Lee, H., Chai, J.: Cyclenet: Rethinking cycle consistency in text-guided diffusion for image manipulation. Advances in Neural Information Processing Systems 36 (2024)

    2024

  19. [27]

    Medical Image Analysis80, 102520 (2022)

    Zhang, R., Cao, Y., Li, Y., Liu, Z., Wang, J., He, J., Zhang, C., Sui, X., Zhang, P., Cui, L., et al.: Mvfstain: multiple virtual functional stain histopathology images generation based on specific domain mapping. Medical Image Analysis80, 102520 (2022)

    2022

  20. [28]

    In: SIG- GRAPH Asia 2018 Technical Briefs, pp

    Zhang, Z., Li, M., Yu, J.: On the convergence and mode collapse of gan. In: SIG- GRAPH Asia 2018 Technical Briefs, pp. 1–4 (2018)

    2018

  21. [29]

    Advances in Neural Information Processing Systems 35, 3609–3623 (2022)

    Zhao, M., Bao, F., Li, C., Zhu, J.: Egsde: Unpaired image-to-image translation via energy-guided stochastic differential equations. Advances in Neural Information Processing Systems 35, 3609–3623 (2022)

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.